Make Kernel registed as a function

paddle/fluid/framework/op_registry.h

@@ -91,7 +91,10 @@ struct OpKernelRegistrarFunctor<PlaceType, false, I, KernelTypes...> {
     OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType(),
                      StringToDataLayout(data_layout),
                      StringToLibraryType(library_type));
-    OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KERNEL_TYPE);
+    OperatorWithKernel::AllOpKernels()[op_type][key] =
+        [](const framework::ExecutionContext& ctx) {
+          KERNEL_TYPE().Compute(ctx);
+        };
 
     constexpr auto size = std::tuple_size<std::tuple<KernelTypes...>>::value;
     OpKernelRegistrarFunctor<PlaceType, I + 1 == size, I + 1, KernelTypes...>

paddle/fluid/framework/operator.cc

@@ -651,7 +651,7 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
     dev_ctx = pool.Get(expected_kernel_key.place_);
   }
 
-  kernel_iter->second->Compute(ExecutionContext(*this, exec_scope, *dev_ctx));
+  kernel_iter->second(ExecutionContext(*this, exec_scope, *dev_ctx));
 
   if (!transfered_inplace_vars.empty()) {
     // there is inplace variable has been transfered.

paddle/fluid/framework/operator.h

@@ -347,9 +347,9 @@ class OpKernel : public OpKernelBase {
 
 class OperatorWithKernel : public OperatorBase {
  public:
+  using OpKernelFunc = std::function<void(const ExecutionContext&)>;
   using OpKernelMap =
-      std::unordered_map<OpKernelType, std::unique_ptr<OpKernelBase>,
-                         OpKernelType::Hash>;
+      std::unordered_map<OpKernelType, OpKernelFunc, OpKernelType::Hash>;
 
   OperatorWithKernel(const std::string& type, const VariableNameMap& inputs,
                      const VariableNameMap& outputs, const AttributeMap& attrs)